AIAW Podcast
AIAW Podcast
E41 - Applied AI Research in Practice - Sepideh Pashami
🎙️ Tune in to AIAW Podcast – Episode 141, Live Today! Join us as we dive into the world of Applied AI Research in Practice with Sepideh Pashami, a senior researcher at RISE and Associate Professor at Halmstad University. From her beginnings as a roboticist at Örebro University to impactful collaborations with Volvo and Toyota, Sepideh’s journey showcases the convergence of academia and industry. Explore cutting-edge topics like predictive maintenance, federated learning, and the role of AI in optimizing industries from healthcare to vehicle battery tech. Gain insights into Sweden’s thriving AI ecosystem, the challenges and opportunities in public sector applications, and the ethical considerations shaping AI's future. Don’t miss this inspiring and knowledge-packed episode!
Follow us on youtube: https://www.youtube.com/@aiawpodcast
We're doing a conference where there are some academic papers, there are some industrial cases like Airbus, or someone coming and talking from Siemens or whatever, or Scania, and then we have Dardox is there, who is sort of working in this space but very practical out with the consulting and stuff like that, and then rice.
Sepideh Pashami:so it's it's different types of players, that that you found and those that became a very nice mix yeah, I have been following research that olga is doing on predictive maintenance last couple of years, especially the way that she's combining machine learning technique or using them for solving the predictive maintenance tasks that are available Pretty impressive. And we have been collaborating also on different occasions.
Henrik Göthberg:Yeah, you said something. She has even helped you in some of your PhD.
Sepideh Pashami:Yes, she was a grading committee of one of my PhD students and she that she, my PhD student, graduated last year, so it's not long ago. And yeah, we have been working on transfer learning and federated learning for predictive maintenance and she is very close to what she's doing as well, so we invited her to be a committee member and evaluate the PhD quality in that case.
Henrik Göthberg:And you were saying it before, but let me go back to it. I mean, one of the things that I think is quite nice with this conference and this was the second this year that was sort of trying to mix up to, trying to find a space between academic conference and more only industrials or industry. So how would you, how, what's the benefit of that, or how are you thinking about that?
Sepideh Pashami:I'm working closely with industry generally. In the last couple of years and like, let's say, last decades, I have been working with industry and what is fascinating is there are lots of interesting challenges or research questions that come from the practical questions, and then it will be very challenging to solve them in a very nice way using AI or machine learning. So for me, bridging this gap is super interesting.
Sepideh Pashami:And often you have an idea of developing something new just because you are aware of certain challenges. If you would have worked theoretically, you wouldn't know about that challenge.
Jesper Fredriksson:So you have a wide field of experience that you can borrow from in the new challenges that you see.
Sepideh Pashami:Exactly so. It's like one of my colleagues says we have a toolbox of methods and so on, and then you try to see which one will fit here, and then how should I approach this problem? But then, if I approach it this way, then how can I handle?
Henrik Göthberg:I don't know that speed of that they need or like that quality or of the result they need, and so on.
Sepideh Pashami:So it's like you try to shape it.
Henrik Göthberg:But is this a common profile in terms of meetups and conferences which is mixed like this? Because I don't think it's. I've been to one more John Bosch at Software Center in Chalmers. They've had around software product management and digital product week. They've been doing something together with Malmö University, so that was in Gothenburg in May this year, but then I haven't seen so many. That is sort of right in the middle, like this or where the academic seat is trying to. Yes.
Jesper Fredriksson:It's one or the other. You're very deep into the research, the base research, or you're just doing engineering? Yeah, exactly this is some kind of middle ground.
Sepideh Pashami:Yeah, this is the middle ground, and usually if you go to one of the two end and you're not for the same population, you get bored of that or kind of. So it's good to have something in middle everyone then. I I would say it needs to be topic based, otherwise again it loses its focus. I think what was good about this conference was that was about one specific topic called predictive maintenance, and that's kind of and I guess that's the logic also what John Bosch did, because they talk about digital product management.
Henrik Göthberg:So the topic base is really around software center. So it's also topic based from another angle.
Sepideh Pashami:Yes, Because this bigger conferences like NIPS, icml, ecml and so on, you see also like people from industry come there, but often just from the research body of the bigger companies they are present. So it's like that's a different population that right now we are talking about, I suppose.
Henrik Göthberg:Yeah, and we of course are used to sort of the industry conferences with the Data Innovation Summit, the NDSML, and I mean like so there's a distinction here between data innovation is more broad and if you go to NDSML it is becoming a little bit closer, more hardcore, technical, in terms of machine learning and stuff like that. So it's like a spectrum here, I guess. But I think this applied AI research topic and finding a conference setting around that is quite interesting.
Jesper Fredriksson:Should we go back to? Who are you?
Henrik Göthberg:Yeah, I think with that, that was the segue, wasn't it? Yeah?
Sepideh Pashami:As you know, I'm Sepideh Pachavi. I'm a senior researcher at the RISE Research Institute of Sweden. I'm also associate professor at Halmstad University. I started my research life when I joined Oreblu University for my PhD. At that time, I was a roboticist. Oh nice, I was not a AI person.
Jesper Fredriksson:Coming back into fashion now.
Sepideh Pashami:Yes, but then toward the end of my PhD I felt more and more that I enjoy this data analysis part of the things. So I enjoy more and more machine learning part. So, then I decided for my postdoc to go to a machine learning group and then I had the possibility of doing this applied research together with Volvo. That was the initial projects that I had that I started working with there. I even went to the company like one day a week and yeah that I learned a lot.
Henrik Göthberg:Was this before RISE? This was before RISE. This is before RISE yeah.
Sepideh Pashami:Yeah, and then after a while when I learned more and more, so my academic career kind of progressed quite a bit. Then I moved to RISE. I joined RISE. I didn't move fully, sorry. I joined RISE, I still kept the Amsterdam.
Henrik Göthberg:You kept the academic foot. Yes, exactly.
Sepideh Pashami:I mean partly because it gives me a very me this possibility of staying and knowing always the high-end research, because of the academic food and because of working with PhD students, and then at the same time, when I work with Rice, I work with a larger pool of companies and also public sectors that potentially are interested in applying. Ai so it gave me a good competencies like complimentary, at least for me so far.
Jesper Fredriksson:Yeah, when did you realize that you wanted to be a researcher?
Sepideh Pashami:I actually didn't know. To be honest, after my PhD I was considering joining industry. Yeah, after my PhD I was considering joining industry and so it didn't matter if I go to the industry or I stay in academia, to be honest at that point. But then suddenly I got the offer from the academia much faster and then when I joined and I did more of the things I liked, so I felt like I can. I can?
Sepideh Pashami:I can make a change, yeah, exactly yeah, I can enjoy it and and, to be honest, also, another thing that is a little bit describes me is I have a passion for teaching. When I see somebody growing as if, like I get extra energy or something, and I, when I see my students, they're going, they're very successful after their graduations or their thesis or their then I feel really well, that's kind of a very motivational SL Just one minute on the Hamstah University associate professor role.
Henrik Göthberg:so here you have. What does that mean in practice?
Sepideh Pashami:so I have a couple of PhD students, so I used to have four or five, now I have three, I think. At Hamstead University I also teach a bit, not a lot because of time?
Sepideh Pashami:Yes, exactly, but I love teaching. I have one course in master level and then one course in education for professionals, which is more theoretical, and then I also have some projects there, to be honest. So one of my biggest projects is called Keeper in Halmstad and it's a collaboration between Volvo, Alfa Laval and Toyota Meteor Handling and a company called HMS which is for networking. So it's a long-term project and I deal with a lot of people, so around 50 people are involved in that project.
Jesper Fredriksson:That's a big project. It's a big one yes. How can you choose between the different? Can you do basically anything you want, or do you have to teach a certain amount? Or how does it work?
Sepideh Pashami:I have been lucky enough to have the flexibility of choosing the things. I would like to do In general. If you can kind of get research project, then you have the possibility of adjusting that how much teaching you want to do.
Henrik Göthberg:But that depends and then we also need to acknowledge we are now here in the ninth season and sahar, a good friend of you, was here in the first season and I didn't know until you told me that you knew each other even before the PhD work. Or you knew each other from way back.
Sepideh Pashami:Yes, I mean Sahar is, like I can say, she's almost a family to me, more than a friend. Yes, we have been having a similar journey, at least starting at the start, like we had a similar education during our bachelor master and then also similar during our phd, and then I decided to move to industry and I stayed in academia, so we diverged at that point yeah, but even up to the robotics and robocop you were, you were both doing that, I guess.
Sepideh Pashami:Exactly so. We were both in the RoboCup competition in Japan in 2005. And we had a team for a soccer simulation. It feels like ages ago, but it was a nice exercise a fun exercise. I don't remember, maybe for the listeners RoboCup, is that a soccer game that you robocop competition is, uh, the idea of that competition is that, uh, they would like to enable robotic to reach a level that the robots can play, basically soccer with human, and so they would like to make it both from physical shape and also intelligent and communication.
Sepideh Pashami:And there are several research questions that come to the picture to make that happen.
Henrik Göthberg:So it is the World Cup for robots playing football soccer.
Sepideh Pashami:Basically yes, usually like there is also a competition between humans and robots.
Henrik Göthberg:Also.
Sepideh Pashami:Yes, just to show.
Henrik Göthberg:But, as someone said, the goal is to beat the humans before 2050, or something like this.
Sepideh Pashami:Yes, that's true. That was the goal. I don't know if they have changed it, but it was written like that that by 2050 we would like to reach the level that robots can play in front of the human.
Jesper Fredriksson:And what do you think? Do you think we will reach it before now? It seems like there's so much progress also in robotics.
Sepideh Pashami:Yes, I think I don't know if we will be reaching 2050. It's too close, but we have progressed quite a bit. Like if you look at the generation before. Like 20 years ago they barely could even shoot the ball.
Henrik Göthberg:And nowadays it's kind of a different story, and the point with this is there are also different leagues.
Sepideh Pashami:So, like the full-grown robots, the smaller robots and the simulation, the simulation, yes, and because of the different leagues they also have different research focus. So, for example, a lot of questions like multi-agent simulation and so on comes to the picture when we talk about simulation. But then when it comes to the humanoid robots, then you have to work with kinematic and optimization quite a bit to optimize the movement, especially when they are trying to shoot, then they might lose the balance, and so on, so it's like there are different research questions come to the picture in different levels, or, for example, the robots that are small size.
Sepideh Pashami:the planning or how they play, the strategy of how to play the game becomes important. So it's like different problems.
Henrik Göthberg:Yes, well, with that introduction of you, separe um, let's move into the theme of this podcast, um today and we we casted it as applied ai research in practice. So it's it's an interesting. The starting point is we went to a conference on this topic, but it kind of brings us down to what we're going to talk a lot about today's applied AI research in topic, and I'm going to set up the scene and then I'm going to start, you know, unpacking it with you. But the reason we chose this is first of all, the reason we chose this is, first of all, you're working at RISE and RISE is potentially an untapped resource for Swedish public sector and enterprise to do applied AI research in a very nice way. So we want to learn all about that.
Henrik Göthberg:And then there's the bigger question of how should we think about applied AI research? I mean, like we are moving so fast now we are experimenting left, right and center. So what is engineering, what is research and how to go about it in corporates is an interesting topic. So, with that set up, my first question to you I mean like to start this off how would you unpack the theme? You know, how would you sort of apply AI research in practice? How would you? How would you define? How would you frame that? How would you unpack that?
Sepideh Pashami:You just said an unpacked resource, right, an unpacked resource, and that caught my eye a bit, so Untapped, untapped, yeah sorry.
Henrik Göthberg:Untapped.
Sepideh Pashami:Untapped resource and that was a bit, to be honest. I agree with you. It's not like we don't do AI research. We do a lot of AI research at RISE and especially if you have been like, interface to the European level of what can be done by AI, and we have been developing in several different layers, from the infrastructure to what it needs to do, like in the high-end research. But I have a feeling that it has been all under the radar so not so many people know about it. This is my point.
Henrik Göthberg:My point was not that you're not doing a lot. The point is that when people are thinking about getting help, they don't know about this channel.
Sepideh Pashami:Yes, that's my feeling too, so I agree with you in that regard.
Sepideh Pashami:But generally, rise has been there, or like, strategically placed there to strengthen Sweden, sweden industry, strategically placed there to strengthen Sweden, sweden industry, sweden public sector and government altogether.
Sepideh Pashami:And so and AI research is strategically became important since a couple of years ago, and we have been also like before that we have been focusing a lot more on digitalization, but now the focus has been a lot more and more on AI, how we can use AI in different sectors. Basically. And sometimes when you say, like, how do you start doing applied AI research, sometimes companies come to us they say we have this problem, can you do something to us. They say we have this problem, can you do something? And then we try to find the right people, that they have the certain competence, and try to formulate the problem and try to solve the issue. But sometimes that's the other way around, that we have an interesting problem, we know there is an interesting research going on. So we reach out to some of our contact in industry and say, oh, I would like to investigate this and I think this will be beneficial for you down the road.
Henrik Göthberg:So there really is. Sometimes someone has a problem and they think, oh, we kind of need research applied to this to start understanding it better. This is when they come to you and sometimes it's a little bit like oh guys, you haven't waken up to the importance of federated machine learning, so maybe you should have a look at it, depending on how you probably have an ecosystem of data you need to deal with.
Sepideh Pashami:And then but a lot of times also we get a question that we want to do some AI, but we don't know what. And that's kind of then we try to set up workshops, try to understand their problems, try to understand which kind of data they have and so on.
Jesper Fredriksson:It sounds like my job.
Henrik Göthberg:And a lot of times people only think about we go to the consultants to get started with this or something like that, but you are a senior researcher at RISE. How would you explain that role and how is that role organized within RISE?
Sepideh Pashami:Generally, rise is a very flat organization, so the titles are not very strongly designed. We have researchers and senior researchers and then some people in the management, so it's like you can tend it. Some people have slightly different, but it's it's very flexible. It's not like you you have the possibility of deciding what you want to call.
Sepideh Pashami:I guess, but so most of the time, as a senior researcher, I'm responsible for making sure that we can deliver what we are promising. So so the responsibility would be more on the senior researcher than the researcher, let's say can you always do that?
Jesper Fredriksson:I'm still trying to get back to this what is applied ai research, and there's like often if it's research, then you more or less don't know if you can solve the problem. That's at least my experience. And then when I come in, that's like I have a pretty good feeling that this can, this can be solved yes, because you you call yourself an ai engineer, Jesper, but you're doing a hell of a lot of experimentation still.
Jesper Fredriksson:Yes and I think everybody is doing that more or less today because it's relatively easy to experiment, and would you call that applied research? Or when does it become applied research?
Sepideh Pashami:So I think at least this is how I see the difference between the consultancy and the research end. So when you have a clear task, you also have a clear solution for it then you don't need us, so you will give it to the consultant and they can do it.
Henrik Göthberg:And they can do it even production level, let's say.
Sepideh Pashami:But what we do instead is like we create prototypes. That potentially shows that, how well we are doing something yeah and that's kind of, that's a research shape of it but still because of our experience in different working with different data, different industry and so on. So we have a feeling that potentially this solution could work if not, we go like back or we try this slightly different way so we have backup plans and so on.
Sepideh Pashami:Even for research you get a feeling of that could potentially work or not, like that's kind of part of the job I would say as a supervisor sometimes I think I understand sort of the distinction I'm I'm often finding myself employing master's students.
Jesper Fredriksson:If I'm like this, I'm not sure that it will work out and it's not mission critical exactly. But then if, if it becomes like too difficult for a master's student, that's many.
Sepideh Pashami:That's when I should call rice. Yes, that would be good, but could it also be.
Henrik Göthberg:I'm trying and testing another angle on on the difference here, because a lot of times, even if you're a consultant or even if you're in in industry, we say we need to do a pilot and a pock of something. And maybe also there's a difference and distinction here that if you go to a company like rice or an organization like Institute like Rice, so instead of the consultant doing sort of a POC in a, I would argue, unscientific way, you could argue with a more rigor in sort of a research approach, phd approach and coming from that angle, it potentially sometimes, when it's difficult, you need more of a scientific method in how you do the, the pilot and the pock and stuff like that would that?
Henrik Göthberg:be a fair uh distinction as well. Like consultants do pocks, what's the difference when you do a proof of concept?
Sepideh Pashami:yeah, we also do proof of concept and we have to do it like that's the nature of ai. I would say, doesn't matter who does it like in the company, we have to do it. Like that's the nature of AI. I would say it doesn't matter who does it Like in the company, they have to do it In the consult, they have to do it and we have to do it too.
Sepideh Pashami:You have to get a feeling about the data, and that's the only way of getting that. I would say that the difference is like sometimes we try technologies that are not commonly used, and that's maybe the main difference. At least we hope so, that that's the main difference.
Henrik Göthberg:I think this is a very good angle and we can take the example of federated machine learning and we have good friends all of us with the scale-out guys and this is still a technique that is maturing. It's not commercially available in the large vendors yet and stuff like this. So this is also a very good angle, I think, where you fit in Exactly.
Sepideh Pashami:So, for example, let me tell you a good example of this. So if you use federated learning in the hospital setting, you often use it because of the privacy reasons, so you want not to share data and so on. That's why you are doing this. But if you do it together with some vehicle companies, so then you don't do it for the same reason, because certain signal might be anonymized enough. Nobody knows what that signal is about. So often you do it because you want to decrease the transmission of the data, because that's costly.
Sepideh Pashami:so then most of the existing literature is trying to kind of not share the data and that focuses on that, or like what's the minimum communication which is actually the problem is privacy preserving data.
Henrik Göthberg:You're not worried about privacy preserving, you're worried about transmission costs.
Sepideh Pashami:Costs, yeah, exactly, and then as soon as this objective change you can. You can change certain things to the existing models to adjust or allow certain variation to the technology that exists and there it becomes a research.
Jesper Fredriksson:Otherwise I, like I'm I'm struggling with the transmission cost. Why is transmission?
Sepideh Pashami:so, for example, you have thousands of features and you don't want to convey all of those in telematics. You can, of course, capture them and and whenever you're connected to wireless or something, then you offload it. But if this is like it's important for something that is safety critical, you want to have that thousand features through telematics and that's costly, you don't want to do that. So then you can do this kind of models. But this is sort of future I would say I'm imagining.
Henrik Göthberg:Because I think at EPFL you presented I can't remember you presented a project around fleets of vehicles.
Sepideh Pashami:Yes, self-monitoring.
Henrik Göthberg:Self-monitoring vehicles. So this is the example with the Volvo trucks or the Scania trucks, and they have been connected for years, they've been doing telematics for years, but what have they really been sending and how have they been sending it?
Sepideh Pashami:And at some point also you don't want to send more.
Jesper Fredriksson:It just becomes too costly for the customer, so you don't want to, but a federated approach applies, that there's some kind of taking things together from parts and making something out of the parts.
Sepideh Pashami:So basically, it tries to send the parameters of the model rather than the signal itself.
Jesper Fredriksson:So it's in this way compressed, so you compress the data.
Sepideh Pashami:Information. Yeah, it's a way to compress. Yeah, that's the idea.
Henrik Göthberg:But could we be a little bit pragmatic or practical. So could you kind of take us a little bit through. Let's put up a case like a scenario. I have a question that maybe has a bit of research in it. I'm working in a medium-sized company and I'm thinking maybe I should try to ask the question or raise the problem to Rise rather than to my local consultants. What's the experience? What is happening here? You raise the question how practically does it work? Or how do raise a question how practically does it work? Or how do you engage with RISE? How does it work?
Sepideh Pashami:So RISE often is a very bottom-up organization and so you often find somebody at RISE and then reach out to that one contact that you have. Feel free to contact me.
Jesper Fredriksson:That was nice.
Sepideh Pashami:No, it's just joking. No, no, no, I'm quite busy to contact me. That was nice. No, it's just joking.
Henrik Göthberg:No, no, no.
Sepideh Pashami:I'm quite busy, to be honest, but then we will, internally within RISE, try to find the right person that has the time and the competence that solve the problem. Another thing I like about RISE and makes it a bit unique is a range of competences that exist. For example, we have interdisciplinary projects from AI and material science and we have all the expertise we need in-house.
Henrik Göthberg:Yeah, because you have so many institutions under the umbrella of RISE.
Sepideh Pashami:Yeah, exactly so if sometimes a combination of the researchers from different departments can solve the problem, not the AI person alone, so those cases are also very interesting.
Jesper Fredriksson:So does that work? Is everybody sort of within the same roof, or is it more like? You have some way of communication that you apply.
Sepideh Pashami:It's often online communication. So, there's 30 different locations, 3,000 different people. But, for example, I'm working together on a project called AI SACS. It's a collaboration between RISE, AstraZeneca and Tetra Pak and Max4 in Lund, and that project was very far from my expertise, so it needed a lot of background knowledge about this SACS curves and imaging of the certain material, and so in that project we are like several people from different departments and then while we are in the same discussion or meeting, then we will figure it out what we can do?
Sepideh Pashami:if we are alone, almost we cannot solve it. So this is quite interesting project in that regard.
Jesper Fredriksson:But if you need somebody like you, you have the feeling that there's somebody in the network, but you don't know which one. How do you find it?
Sepideh Pashami:Usually then you will ask your manager and then it goes down the management level if you really need somebody and you have no idea. But after a while being in the organization you get a feeling of who you can ask, maybe you should have a Slack bot that knows everybody so you can ask the slack bot we have also something called ai center internally at rise yes, yes and uh he's really nice and, uh, what he has done is like uh, he created a ambassador group, uh, 70 different ambassador group that they are in contact with other parts and then they will be in contact with us.
Henrik Göthberg:It's like a core AI researchers in the AI center so that's another way of communicating, but it was actually one of my private curiosities or questions, because we had Sverker Jansson here on the pod and we understood about the AI center. And how are you now? You're sitting in one of the legs or one of the hubs, so I don't know how you call it yeah and you are AI researcher. So you you have like a line manager or like this, and then Sverker's center is going more across. Could you elaborate? How does it work?
Sepideh Pashami:so kind of. So my, I have a line manager that I kind of responsible for what are the projects I'm working on and then how, who I'm working with or which resources I need, that she, she helps me to kind of work like find the right resource and so on that I can, because many of the projects that I'm working on I'm not doing them alone, like no yeah, almost you're the senior researcher and then you have researchers exactly so it's like, therefore, I really need her help to kind of give me the resources I need, given the and also our projects are time limited, resource limited and so on.
Sepideh Pashami:So it's like we have to find the right person with the limits that we have, that we can deliver kind of what we need to deliver, and then that's kind of responsibility of hair. But then because a lot of our projects are research-based, research-heavy based, then we have sometimes little time to do theoretical research or learn from each other or talk to other senior researchers at RISE. That would be the role of the AI center that we are kind of like in the. We are kind of AI core team. We have AI core team that we discuss in a certain strategical point that this is what we would like to do during this year. This would be like a new topic that we want to learn, and often it's not just we want to learn, we also want to teach to the organization or use in different projects. So then those come from the AI Center group. They are not necessarily bonded to a project.
Henrik Göthberg:And when we talked, it's quite obvious that I'm not sure I need to ask you if this is normal, if it's you, Because sometimes we do projects and the project you do one project and then it ends. But when we look at your profile, your projects never end. So they end, but they iterate. So we were joking about this, that you have an exponential growth in work because you have one project and then that continues, and then you have another project and that starts. Is that how it works? Are you out of the norm here?
Sepideh Pashami:I guess I'm out of the norm a bit here. Yes, I'm a very curious person, I would like to explore and I'm not afraid of learning new things and doing things slightly differently.
Henrik Göthberg:So you see problems and then, when you solve that problem, you see the next problem, and then you tell them about that exactly so.
Sepideh Pashami:That's how it works so I I would say that's that shows actually that they were happy with the collaboration yeah, and in this regard also, I think the collaboration is a complex uh setting. Uh, it's not only about like doing the job. I think, like in the initial collaboration I had, sometimes I would even hold the lecture for the team that I was working, just to have the same language yeah, this is super interesting how you start when you don't have the same language, exactly I would.
Sepideh Pashami:I'm a volunteer like I, going to give you a lecture on this.
Jesper Fredriksson:And they eat it up. They eat it.
Sepideh Pashami:Then after that we will be able to talk. You know, they will understand why I'm saying that I need X data or that's not enough, or I need more data. Then they will understand what it goes in and that's kind of missing.
Henrik Göthberg:But this about collaboration is such an interesting topic, but let's be really concrete. So let's be a little bit like we are a fly on the wall and we want to understand how does the collaboration with the Olympic Committee work compared to how does the collaboration with Volvo work and how does the collaboration with the AstraZeneca Tetra Pak? You know, if you contrast, are they working the same or very differently?
Sepideh Pashami:They are slightly different from each other. So I mean I have been working with Volvo group the longest, so last continuously last nine years I have been working with Volvo Group for the longest, so continuously. Last nine years I have been working with them. A different part of Volvo Group and what I love there in that collaboration is now there is some sort of a mutual trust between us.
Henrik Göthberg:You work so long, so you know what you have each other.
Sepideh Pashami:Exactly so. It's kind of you feel like I can count on this. It's a different level that I don't need to prove myself every time that I meet them again. But and has been also reaching, let's say, closer to production level of research. Just because it has been long.
Sepideh Pashami:And that's kind of more than just experimentation, and it's like now we are working on proof of concept, proof of concept on vehicles. So it's kind of finally reached the point that we can be proud that what we are doing exactly, or? Research will be implemented, so that's kind of hopefully.
Sepideh Pashami:So at least we are showing the proof of concepts, and then it will be on the company to decide if they want to use it or not. That's not on us, but still it's kind of exciting to see the impact of your work. But then what I liked about collaboration with Swedish Sailing Olympic Committee was it was a very concrete task collaboration with Swedish sailing Olympic committee was. It was a very concrete task, but then with high impact gold and medals exactly.
Sepideh Pashami:I had a feeling that the impact is so close. If I do a good job, then I can see the impact next year tell the story, tell the story.
Jesper Fredriksson:So we started a project with a, so then I can see the impact next year. Tell the story about that.
Sepideh Pashami:Tell the story. So we started a project with a colleague of mine Her name is Laura and she's in the maritime department of RISE. She actually reached out and she herself used to be a sailor and she has been collaborating with the Olympic Committee and she came to us and said that okay, I need some AI people to help us, because every small improvement matters in this competition and we were like, okay, we think we can do this, we can do that.
Sepideh Pashami:We tested here and there different things and at the end we decided that, okay, something that seems beneficial for them. At the same time, easy to do with a short amount of time we had, was looking into the strategy of movement of the sailors during the competition, given the wind information.
Henrik Göthberg:With the wind information optimizing strategies for sailing, so with the wind information optimizing strategies for sailing.
Sepideh Pashami:So we wanted to optimize the strategy and we knew that the wind pattern might change. So we had several wind patterns and then we would try to predict the wind pattern. And on top of that kind of estimate, the stage. How did you know that that was a good angle? I mean, it was important for them, that was one thing. And then I also had an idea that how can I solve this in a short time?
Sepideh Pashami:So I thought that could be something, but to be honest you're never sure, so on the side I was doing also small other things that to see.
Henrik Göthberg:So you had alternative ideas of strategies. How could, could use AI. But here you had a fit, a problem that they needed to solve and actually a pragmatic solution.
Jesper Fredriksson:Sorry, I was just thinking about the process. It's fascinating because I see this all the time. You see a problem that you need to solve and then you need to talk to people and find out, because you're not an expert, I guess that you need to solve and then you need to talk to people and find out, because, yeah, we had to find the data actually like like we went back.
Sepideh Pashami:So laura went back and talked to them and said, like, which kind of data you can give us? And then they talked to some coaches and then they said, okay, we can get some data from some of our training sessions. We got those and then we felt like no, that's not what we need. And then we looked into the previous, like competitions and so on and then we tried to grab those ones and then and then we felt like these are better like then we were not sure.
Sepideh Pashami:You see why we were not sure? Because we didn't have all the information, and after that we kind of okay most probably we can be able to solve it.
Henrik Göthberg:But I love your question because this is such a applied AI research. To me, it's like you're trying to get your hands on materia where you have unknowns, unknowns, and you need to start somewhere. How do you do that process, you think, because this is one of your, I think, golden skills or superpowers.
Sepideh Pashami:Yeah, we have to learn a little bit about the problem. And so sometimes I visit the company, I go in person and talk to experts. That helps a lot, and then what's possible to collect and not, and so on.
Henrik Göthberg:So it's like two major pieces Collect the problem and collect the possible data you could think about. Those are two major strategies.
Sepideh Pashami:Yes, yeah, yeah, exactly. So that's kind of. Sometimes also, there is lots of data but not what you need, kind of. So it's not always having a lot of data will definitely solve the issue. Sometimes, at the end, you need to talk to experts and find patterns or incorporate some of their knowledge into the models that you'll be able to solve it with a lot less data, and so on. So it's not one solution fit all at all.
Henrik Göthberg:Is this a common theme in all projects, that this is how it works?
Sepideh Pashami:I mean, we have all kinds of projects, so some that they come to us and they don't know what to do. Then we often with those like we say, okay, these are the possibilities based on what has been happened in the last year and what we have heard of and what we think it's possible to do as well. And then the next question always is data. Do we have data? Do we have ethical permissions, considerations that we can use that data, and so on. So, for example, that comes to the picture when working in the healthcare domain yeah, exactly A lot more than the others. And then we start to formulate a problem. We often try to choose one low-hanging fruit first, just to get to the data and then deliver something that's kind of we know, it's kind of doable, and then from there we also define that okay, now we want to explore these things that they sound crazy.
Henrik Göthberg:If they pay off, they pay off really good, so it's kind of you always try to balance that.
Sepideh Pashami:At least this is how I'm doing it. I guess not everybody does it.
Henrik Göthberg:But when you're working with the Olympic Committee, with the sailing, when was this and what was the result? We need to take the whole story home, because it's a beautiful story.
Sepideh Pashami:Yes, so we started by looking at the data. It was me and another colleague, bjorn, and then we hired a bachelor's student that kind of helped us to do this and his name is federico. And then federico was asking me, like he was looking into data, telling me what's possible, how to clean it, what are the features we have, and so on, so forth. And then at the same time, also b Bjorn was looking at the data and I remember for like a month or two I had no idea what we wanted to do. I was not telling them, obviously, so they were looking up to me, but I had no idea.
Henrik Göthberg:You played it cool for two months.
Sepideh Pashami:Yes, but because there were a lot of challenges with data, so we had to deal with those first anyway. So I had time to think and then I came up with an idea that let's do this method and then let's do this method still meant we couldn't use it off the shelf. So because, for example, we use a method called ASTAR it's a very old, traditional model for optimization and especially for gaming to find the shortest path. But often on those problems you're dealing with a kind of a limited space and also a very discrete kind of environment. But our environment was not discrete, and then it was not. We couldn't use it just blindly. So then we started to think how we can do that function, how we can work with that function, and then how we can put on the angle.
Jesper Fredriksson:Did you make it discrete?
Sepideh Pashami:no, we were using the angles, and then we just keep track of the angles that we have explored, so we didn't make it discrete at the end, using the angles, and then we just keep track of the angles that we have explored. So we didn't make it discreet at the end. So we did slightly differently and we published that work also. So it was not only we helped, kind of the Swedish Olympic Committee.
Henrik Göthberg:And when was this? Which Olympic game were you?
Sepideh Pashami:This year, yes, 2024.
Henrik Göthberg:So it was coming up to 2024. Yes, and I was watching this on television, me too, I was heavily following.
Sepideh Pashami:And how did it go? It went really well. I mean, the sailors in the Olympic team, they were amazing, to be honest, and I was so proud and I was happy to see that I had a small part.
Henrik Göthberg:So how many medals did we get? Two, one bronze and one silver and was it working like this, like you supplied now a toolbox for the coaches, for all the coaches of all the Olympic sailors.
Sepideh Pashami:Or was it focused on some classes Four classes, four classes, four classes that they were competing in the competition? But of course we would like to do more for the next Olympics.
Henrik Göthberg:But four classes, and within those four classes you had two medals.
Sepideh Pashami:Yes, exactly, it was not honest.
Jesper Fredriksson:I mean I would-.
Henrik Göthberg:I think it's your medal.
Sepideh Pashami:I'll give it to you, no way, it's the hard work of the sailors, for sure.
Henrik Göthberg:We were there to help, but it was fantastically fun. I can imagine when you've been working on the model. I know what you told your father. Imagine, oh, I get goosebumps. It's so cool.
Sepideh Pashami:Amazing it's good fun yeah it's not the, I mean the whole team. I would say like we were, and now, because of this, there's some more.
Henrik Göthberg:I heard somebody said curling yes, yes.
Sepideh Pashami:What are you planning to? Do for curling yes, because, as I understood, curling is a very challenging game too, so the stones are different from each other, and then also, like during the training, the consistency of the players matter, and also, like for a particular location and and the consistency is not obvious to do like, a small angle difference from the way you are throwing will lead the stone to go far away.
Henrik Göthberg:So, once again, data-driven coaching of how to improve your game.
Sepideh Pashami:So we talked actually with the coach of the women team of Keling Olympic and he suggested that what he would like to see is if we can suggest if there is a slight difference in angle when throwing the stone during the training sessions. And another thing that they were interested in was like looking into the differences with. Can we highlight the differences between different stones in the competition, because before the competition starts they have a little bit of time that they can test the stones.
Henrik Göthberg:Test the stones and select the stones for the eyes.
Sepideh Pashami:Yes, select the stones which stone is good for which kind of shot. That's kind of what they are interested in, but we are just starting this project.
Henrik Göthberg:But let's solve the problem here now. Would you need to have video imaging or what different parts to figure this out?
Sepideh Pashami:Right now we are talking with people from the sensor department, so we are testing the cameras to kind of look into the posture of the people at the time of throw. We are looking at the imu sensors, obviously, and then also we have been looking and we have been thinking of some other new sensors like audio sensors and so on, and then we have. We are going to have a data collection session in january, so then then we will will get started, so I was going to ask this question before you asked it.
Jesper Fredriksson:But it's beautiful. So you don't even have the data yet, but you have to devise a data collection scheme first. Yes, wow.
Sepideh Pashami:In this case. Yeah, that's what I was saying, like that's the difference maybe between the applications.
Henrik Göthberg:This is research.
Sepideh Pashami:Yes, but for example, when I work with a Volvo group, then we have the data, so it's like we have to work with the data that's already collected.
Henrik Göthberg:But this is another interesting thing when you have an institute like RISE to piggyback on, because when you start breaking down the problem it moves very quickly out of the algorithm model problem into hardware sensor problem, imaging problem, all this right. So all of a sudden, like when you're going into the real applied research and you're doing things that are sort of physical, it is quite hard for any consultancy to master all that.
Sepideh Pashami:Yes, exactly that's why that's kind of especially interdisciplinary works. That needs different competences.
Henrik Göthberg:That's kind of so, if we were sort of kind of find a sweet spot for when RISE is really fantastic partner. I think it's when it's like these multidisciplinary problems like material AI you know that seems like who can do that. It's not easy, Even in sailing project.
Sepideh Pashami:It was not just that we look at the. It was not only our AI department that was involved. They were like maritime department that they were leading the project. And also there were another department called additive manufacturing.
Henrik Göthberg:Additive manufacturing even to the materials.
Sepideh Pashami:They were looking at the foilings for the boats. So much math here. Yeah, it's like that. The. Yeah, it's like we were doing different pieces of the puzzle.
Henrik Göthberg:So RISE was supporting not only from the AI perspective. So RISE was supporting the Olympic Committee and from the AI department, from the materials department and from the maritime department, from the materials department and from the maritime department.
Sepideh Pashami:So it was like that. I think that was a very successful project.
Henrik Göthberg:But it's actually. It's a showcase of what is. When should you think about rice? Ultimately and I think that goes a little bit like how we do now you know Scania and Volvo and all them it's like very applied in the terms of its AI, but it's in predictive maintenance or it's in a very physical environment where you might need different competences.
Sepideh Pashami:For example, one of the areas that I'm interested to look into, because of my collaboration with vehicle companies, is looking to the state of the health of the batteries on board of the vehicles, and so I'm looking at it from data perspective, of course. But then we have experts that they are familiar with how the battery actually works. They have the test equipment that they can look into the batteries.
Henrik Göthberg:I need to connect you with my old colleague, Georg, who worked in my team in Scania. He's now running the analytics department for batteries we should talk.
Sepideh Pashami:We should talk about another scenario that again goes cross yeah, exactly, because then?
Jesper Fredriksson:yeah, super interesting. Uh, um, maybe you can teach me something about batteries. I mean, I'm from volvo cars but I don't know anything about batteries. I'm not, I'm not part of the manufacturing of the car, but in my impression it's like if I talk to somebody in gothenburg, they will know everything there is to know, but you're saying that it's still unsolved research questions around it, especially for batteries it's like a new field and it's like a vehicle.
Sepideh Pashami:Companies have been very much familiar with a mechanical component and this is slightly different and we are seeing also a difference in behavior in different conditions yeah, yeah, yeah, definitely.
Jesper Fredriksson:In winter time it's totally different, of course yeah, and that's that's expected. Yeah, yeah that's expected, but I I was. I was under the impression that this is hard, but it's probably like almost solved.
Henrik Göthberg:That was my thinking first, second, third generation, where they're fundamentally, we are replacing some of the raw materials for new raw materials and then when you do that, you get other characteristics of this battery so this is a never-ending story to me and also like the question can go beyond the vehicle industry, so questions like second life of the battery for sustainability.
Henrik Göthberg:I have one of my friends in Scania who's sort of moving into sort of ventures. There is a big game going on now on circular economy. So if we have a battery in the truck here, where does? Now? It can't really go in a truck anymore when could it go now? Could it go in a battery wall? How would that business model work? So this is sort of, is it? It's business modeling research as well? Yeah, do you do that? The rise to business model research? Because it is.
Sepideh Pashami:no one has thought we are figuring stuff out yeah, we don't do it, at least not our department. We are more on the yes, exactly so. But our partners often think in that way, like when I talk to Volvo colleagues or Scania colleagues they think that way? Yeah, they have to.
Henrik Göthberg:And I'm going to switch gears, because now we talked applied AI research very much with a rice focus, and I want to really move into the second dimension, applied AI research in general and how corporates and public sectors should think about it. But before I go there, I think we should have a small AI news section right now, because the segue to do it right now would be perfect, and then we move over to the next topic. It's time for AI News brought to you by ai aw podcast. So in the middle of the podcast, we are trying to pick out some interesting ai news we found. Sometimes there's one, sometimes there's many. I have one major news piece I want to talk about, but I will start with you. Do you know? Maybe this was coming or you didn't know?
Henrik Göthberg:you told me that this might come do you have something you pick up in the news that you want to talk about? If not, it's ok.
Sepideh Pashami:I didn't pick up in the news, but I read a paper but that's fantastic yesterday, let's go there or the day before that I was enjoying Please.
Henrik Göthberg:Maybe I can explain that.
Sepideh Pashami:So the paper is not new. It's 2022 and the paper is. We know. We have been seeing success stories of natural languages, for example in Chachapiti, and then they can kind of memorize a long history and therefore they can kind of generate a new text that kind of becomes human-like text. So the question is can we use the same technology for time series?
Henrik Göthberg:What would?
Sepideh Pashami:be the challenges and what makes it easy.
Sepideh Pashami:And this is kind of one of the research areas that I'm interested in and I have been reading about, and one of the papers that I liked was this informant paper. It's not super new. There are much more newer work on this but it's kind of have a good idea, let's say, on the way. And the idea is that when it comes to time series you often have, you're interested in looking at the longer history, not just the few hours before. So how to encode that? And if they use the vanilla kind of canonical transformer come from the languages then this becomes computationally very expensive. It does not become doable. So their idea is that how can we enforce sparsity during this attention mechanism, that we only keep the important information during this encoding and decoding structure, that we be able to still capture the long-term patterns that we are?
Jesper Fredriksson:interested in and how is the sparsity enforced?
Sepideh Pashami:patterns that we are interested in and how is the sparsity enforced? So, uh, it was some sort of approximation plus some sort of looking into importance in attention mechanism. So in attention mechanism, often, uh, you create a code per word that is kind of embedding per word. That is like, um, for some of the words it's very important and some it's very unimportant. So you try to only keep the part of these embeddings that are having larger distribution and they are not very uniform and flat, so they're ones that are kind of interesting.
Jesper Fredriksson:You only keep those. So in this way they are enforcing the sparsity. And how is it trained? Does it work on?
Sepideh Pashami:any time series. In this particular paper it was working for the time series that they also have a sense of a time or like an order of things. It was not like a foundational model for all types so the the result also. They showed it was for prediction of, for example, temperature, yeah, but then they showed that they could predict like 720 days. Time is done. Sorry, instances, I don't remember if it was minute or hour, and so on. 720 instances.
Jesper Fredriksson:So it can work with seasonality and those kind of effects. So would it work on the stock market?
Henrik Göthberg:That's my first question. I was waiting for the applicability.
Sepideh Pashami:Yes, I have to test and I will tell you.
Henrik Göthberg:But why were you interested? What is the applicability?
Sepideh Pashami:Why are you looking into these particular questions?
Henrik Göthberg:Why did?
Sepideh Pashami:it caught your attention predict the temperature of the high voltage cables across Sweden through the air. And we managed to suggest the solution using a convolutional neural network and then predict. But then our prediction was not very far in future and it wouldn't include lots of information from past. So I was hoping that I can find a solution that I can predict a lot longer if you want to extend the prediction window.
Henrik Göthberg:I know what you do, I've seen your excellence in motion cool. Okay. Did you have any new segments you wanted to share? Yes, but I have, I have.
Jesper Fredriksson:So I've been, as you also have, been very interested in agents and there are a lot of different sort of foundational works that's being done. Foundational sounds like it's very heavy on math or something like that, but it's not so much that, it's more like scaffolding scaffolding things how to practically scaffold this so, um, first there was this uh, both of these things are from anthropic, so claude released this computer use.
Jesper Fredriksson:That was a few weeks ago. That was almost scary yeah, that was it's scary to think about into the wild. Uh, because it's scary to think about releasing data into the wild because it's letting the LLM take control of the computer and do basically anything. So that seemed like an interesting thing. For how do we connect agents to do things in the world?
Henrik Göthberg:Properly like practically.
Jesper Fredriksson:And it feels like, yes, that's how I use the computer, so it's like you can click here, you can click there, and then they count the pixels et cetera. So that was interesting, but I think it was last week or a couple of days ago. The model context protocol was released, also from Anthropic, which is a protocol for how to do tool use in LLMs. So tool use is one of the foundations of agents. You're supposed to be able to do things in the world, so that we have been able to do for a long time.
Jesper Fredriksson:But this is an open protocol that they have released and it will be interesting to see, for example, if OpenAI will pick it up or if they will develop their own. But it's a way of connecting different services. It can be local on your computer or on the web somewhere to the llm, so you can just specify what tools do you have available, so it can be like a sql database, it can be your github login, things like that. So the LLM will always have a way to connect to all these resources. So I've been always thinking about how can I get my AI agent to do my job when I sleep.
Henrik Göthberg:When I saw this, I think this is a fairly big deal. And if you contrast that to the current approach where it's very, very fragmented and we are using different protocols and different approaches in every single piece of the puzzle and it was interesting when I heard your presentation as NDSML Okay, it's one thing to do a simple scaffolding, but if you have an agentic workflow objective, like you to do a product manager, data analyst agent so you have a product manager in Volvo and he basically don't want to hire a business BI guy. He wants to have agents who do the data prep, the structuring and then, ultimately, the presentation. And if you take that on steroids, you need to go out and do different things and even sensor data, different data, different types of protocols this becomes a very, very fragmented, multi-agent model. So having then a universal protocol is a very interesting thought.
Jesper Fredriksson:Yeah, it's almost. It feels almost like.
Henrik Göthberg:Did I get it right, by the way?
Jesper Fredriksson:Yeah, I think so. It feels almost like the HTTP of agents. Yeah, exactly.
Henrik Göthberg:So the HTTP of agents. We coined it now. We said it first. It hasn't been out on the internet. We think about this as the HTTP of agents.
Jesper Fredriksson:Yeah, that was cool and people are doing cool stuff with it, but it's mostly small things.
Henrik Göthberg:I like the way that you said that this is foundational in terms of fundamental protocol that becomes the HTTP protocol of agents. I think it's super cool.
Jesper Fredriksson:It's really interesting. Let's see where it goes and let's see how broad it's. Already a lot of tools that are being put into this protocol.
Henrik Göthberg:But you said it right, it's open. Done by Anth will open. Yeah, I pick it up, or maybe they are forced to you know what's your bet.
Jesper Fredriksson:What's your bet? I wouldn't be able to bet I think it's 50, 50, 50, 50.
Henrik Göthberg:Yeah, because I I can see how they're squeezing yeah, should we do it or not?
Jesper Fredriksson:ah, because, because they have something they added recently that you can connect your coding tools. Your terminal and your notepad can be connected to ChatGPT. So they must have their own way of doing this.
Henrik Göthberg:But I love how you put the HTTP of agents, because if you think about this, like some of these protocols that allowed us to do internet, we kind of need that in order to really scale up agents. Otherwise, you know, this is one of the problems. If you compare telco to energy, right, so telco, they got their 3G, their 4G and they got their you know the consortium and they got the standardizations right. And then in energy, we never managed. They got the standardization right. And then in energy, we never managed and it's a nightmare. Smart meter is here and there, different protocols, lora, la la la, so for LLMs to really scale out, the agent vision.
Jesper Fredriksson:I think this is foundational. We'll see. Maybe it's the XML instead of JSON, I don't know, but I mean json came later and became much more popular than xml.
Henrik Göthberg:maybe maybe this is the xml, I don't know, we don't know if this is the final game. That's true. That's true. Interesting two topics. I'm gonna have one topic as well, and my topic is then a very contrasted topic to you.
Henrik Göthberg:So this week or yesterday, I guess, the AI Commission handed in their report. So the AI Commission was set up by the government and led by Karl-Henrik Svanberg, who put together an expert team to basically, as we do commissions in Sweden to put the report together in order to highlight, you know where should we go with AI in Sweden, and for people who has been on this pod in these nine seasons, we had a few of them from Jyrgini's council, yet it's like, oh, here's another report. But the interesting thing now is that it came out and the first thing that was super cool, this is cool. They got until june 2025 to release the report and I and when I heard, oh my god, you, you're not, you're not figuring this out, this needs to go way faster. And this is the backstory anecdote we had karl-henrik Svanberg on Data Innovation Summit, so I'm the chairman, so you know. So I was sort of, I was shepherding him up on stage and down stage. I had a chance to you know, you chat a little bit and I was like this commission stuff is great, but this is way too slow. I can't remember if I said it. He said it even Like no, no, no, we, we will not use that time, we will. And I I think it's one of the first times ever I've heard about a commission coming in one year earlier or like six months earlier. But he said it. He said it and he said it like henry, call me, I will be on the podcast when we release the commission. So now, now I'm going to chase him down for that. So that's the backstory and the first tick in the box. That, I think, is fantastic.
Henrik Göthberg:They sensed the urgency, they did the report and they released it. Kudos for that, kudos for that attitude. And has anyone read it? I'm reading it, but it's like 136 pages. I'm to page 20, 50. I'm sorry I haven't finished it, and there was an article in Dagens Nyheter like a debattartikel, but overall, my sense is that it's fairly positive in the way people have looked at how the report has been shaped. So they have the sections of like a background and they have, you know, a lot of different stuff that we need to work on. They have some concrete recommendations like 70 of them. They have concrete metrics that we should follow up on.
Henrik Göthberg:But the most interesting recommendation they have of all of them, which I found really cool to put in the report, they are recommending the government to go into what is called in Swedish stabsläge for a certain topic. So it means when something is really urgent like if you have COVID, how you manage and organize and everything should go through all the motions it should take five years. But when something is really urgent or when you're in a catastrophe, you go to stabs leg. I don't know what it's called in English. Can you help me? Emergency mode, emergency mode, right? So they are literally highlighting that for the first couple of years we should put the AI strategy in stabsläge, which kind of means that you're putting a project together or a program together that reports directly to regeringen and that has like a very clear direct agenda to get things done, and then when the urgency is called down, it can sort of sort itself out.
Henrik Göthberg:And I found that extremely refreshing to put this in the report as a recommendation. I thought that was very cool. That's pretty strong. That's a strong statement. Did you pick up on that? Yeah, you saw that.
Jesper Fredriksson:I didn't see that actually, but I picked up on that one. I started reading it from scratch and I didn't see that, but I also read Anders Arptig's the main host. The main host? Yeah, on my seat.
Henrik Göthberg:Arptics, the main host on my seat.
Jesper Fredriksson:I read his comment on LinkedIn.
Henrik Göthberg:And what was Anders' comment? I missed that, was he?
Jesper Fredriksson:happy or was he disillusioned? I think it was both. It was a mixture, and I can subscribe to what he said was the negative part, and let's see if you think the same way, because there was a lot of talk about research and Anders was basically saying that engineering is more important, which is, I mean, that's who I am, so I have to say that. But there is something around realizing where we are on the map. We cannot do a chat GPT. So what should we do? Should we really be doing that kind of research or should we use the tools that are out there already? That that seems like in my world it's, it's a no-brainer, it's, there's some, it's, it's like a gold, it's like a field of gold out now, out now, and there's gold everywhere. You just have to find out, which is not true gold. When you pick it up, you see that, okay, it doesn't work here. But instead of building the mine to get to the gold nuggets, just pick them up and see if they work. That's what I'm thinking.
Sepideh Pashami:There was something that I was also interested in, that there was emphasis on doing ethical AI. That's something that I have been thinking about, especially with regard to robustness of machine learning models and explainability of those. I know it sounds a bit futuristic and especially to companies, it sounds like unnecessary step, but I think we will reach a point that we need to do this. But, on the other hand, like often, this becomes super important to the public sector and government to if they want to use ai.
Sepideh Pashami:They have to understand the limitation, they have to understand what they are gaining and what they are losing in return like, and that's something that we have to work on, and I was happy to see that it was in the report yeah, I love the approach of explainability and ethics.
Jesper Fredriksson:Then it becomes more tangible, exactly, and I, for example, I love anthropics work on that, that, where they try to find the features within the network and try to learn more about what's happening around that, and if we're going to be opinionated, I have sort of two angles of opinions on this work.
Henrik Göthberg:The first one is we always need to do a very detailed analysis in the start of the report. It's very detailed out who has been working on the report, and this is the first level of interest. That ties back to Anders Comet immediately. So you have basically professors in academia, like Per Heinz, you have economists of different kinds, you have economists of different kinds, you have politicians of different kinds and you have ceos of different kinds, like you know, like this ceo of volvo or the ceo of ericsson or the equivalents, but not one hard core like cTO or engineer, and my take on it was a little bit like to be a little bit cheeky if you really want to get good advice on how to talk about data from Scania, should you go to the CEO or should you go to the guy I know, jan Gures, who is leading the AI initiatives in Traton, you know? So it's a huge difference, right?
Henrik Göthberg:And when I look at the profile of the people that has been advising on this, there's the blind spot. So what they have is tick in the box. It's fantastic, but there is a blind spot here and it's the hardcore CTO role and you could flip it into the chief data officer role, or you could flip it into you could go from a CTO like like in one of our hyperscalers at Klarna someone who really, really represents the engineering perspective. This is the same angle. So this is first of all in the report this is when you look at who was in the commission.
Henrik Göthberg:This is first of all in the report. This is when you look at who was in the commission. This is mistake one or blind spot one. If you look in the report and we look at all the different angles, that has been highlighted. The other blind spot actually the one person who explained this blind spot the best was Sverker Jansson, when he was here at the port, because he was contrasting around what it means in Silicon Valley when you have these big tech giants and they're spewing it out, people who really have a fundamental engineering understanding of machine learning and data engineering at scale professionally. And if you look at that talent pool and how that disseminates out into the next startup and the next startup and how that fosters an engineering culture, that is something that we simply don't have right now. There are too few Lalle Lars Albertsons. There are too few Sverker Johns now. There are too few Lalle Lars Albertsons. There are too few Sverker Johnsons. There are too few.
Jesper Fredriksson:Jesper Fredriksson, I do that all the time. You're not alone. Everybody says Fredrik.
Sepideh Pashami:But actually they mentioned that there was a plot toward the end of the report that they were highlighting that Sweden compared to other countries. From talent point of view, they have a smaller pool, but the difference is.
Henrik Göthberg:It's not from the research you get that blood on your shirt. It is from the projects, the startups, the hyperscalers. It's from those you know working on applied AI research and then beyond right Applied AI research and let's go now beyond that into engineering. So do we have an AI research problem or do we have a senior AI engineering problem in Sweden? I think that's the core question.
Sepideh Pashami:I think we have both to be honest, Because you're not in the research end, you might not feel it, but we also need the logic, of course, is like from the research comes, then the people that do the startups and then they trickle out.
Henrik Göthberg:And one of the key comments that they recommended is these combined types of positions part research, part industry. That is kind of going in the right direction what we are talking about now, but I think there is a distinction here again between engineering and research problem. And going back to are we going to do the mine or are we going to do the maximum amount with the nuggets? I think I'm not saying this is bad. I'm just picking up two blind spots and they're connected. They're connected because if you had a cto in there, he might have explained the story differently and then the report would have had another section in it. So I'm not. I think what we have is fantastic. I'm talking about what we not have that might should have been there.
Henrik Göthberg:Anyway, we're happy. Right, it was good Overall. I think it's very positive.
Jesper Fredriksson:Yeah, it's. I mean I like what you said, that stop slag. That's a good state of mind.
Henrik Göthberg:Yeah, let's see if they get it. They were recommending it that they would treat it like that. Mind, yeah, let's see if they get it. They were recommending it that they would treat it like that. Well, let's see if that happens.
Sepideh Pashami:But interestingly enough, I had a chance to talk, to call as well and explain the sailing project oh, he's a sailor of lover yes when did you have that chance? Because when he visited right.
Henrik Göthberg:I was part of the team that we talked to I bet you he was wanting to talk more sailing than anything else.
Sepideh Pashami:That's true, that's great.
Henrik Göthberg:That's awesome, but it's a good anecdote. So any core messages you were trying to convey Was it more nice or was it something that was sort of disgusting? That was quite good.
Sepideh Pashami:I mean at RISE we have different departments. So I'm from analytics department, so data analysis, and also we have intelligent system. That is kind of sister unit to us. But we also have units connected to the kind of creating high-performance computing cloud and so on, so forth and they are working on the recent advanced development there. So it was many different of us like with different expertise in the room that they are all connected to AI, because I mean, if we don't have the infrastructure for training our models, we will not be able to develop our models further than a certain amount, like a certain stage anyway. And then we also had in the room some of my colleagues that they are working in EU level discussing the possibilities of what should be the next regulations level, discussing the possibilities of what should be the next regulations for AI, what should be like the next level of the projects, like consulting on those in the EU level or collaboration potentially between the countries. So that was an or take in it, but I'm guessing they have to summarize everything that everybody suggests.
Henrik Göthberg:So it's like several of those topics that you mentioned now I think they are in the report quite well. I like the EU collaboration, how we need to be careful of how we do regulation. I think that is actually quite well. That is some of the stuff that is in the report in not in a bad way.
Sepideh Pashami:I think it's quite good anyway, and cyber security as well.
Henrik Göthberg:Cyber security is also a big topic and they wanted to build it. We forgot that. One of the key, one other key recommendation is to build some sort of cybersecurity center in Sweden we have one at Rice.
Henrik Göthberg:Yeah, but I guess they want to make it bigger. Anyway, that was the AI news section. I want to now broaden the topic and start talking about applied AI research in enterprise and public sector, and I'm going to start like quite broadly is that relevant? Or and why is it relevant today, you know, do we need to think about applied AI research more in general in companies and private sector, given where we stand in 2025?
Jesper Fredriksson:It's a good question. I mean it's obvious that you have some very good use cases.
Sepideh Pashami:We have also some discussion with Skatteverket and other regions Stockholm.
Jesper Fredriksson:How do you help them? That's super interesting.
Sepideh Pashami:Yeah, so the explainability idea that I expressed is pretty interesting to them. It's because, for example, when they use any machine learning techniques, because they are very close to the end user, so they would like to have explainability. So that was the topic that I have been talking to them and also many of the explainability models that exist. They are often designed to be communicated to the developers, not the end user, but then what they need is to communicate to the end user. So I think there is a place again for doing more research and how to translate what we are often showing to the developer to the end user. That kind of become interesting question.
Henrik Göthberg:And this cost topic goes all the way back to the DIG report that Patrick Ekamo was working a lot about. So how we are able to build credibility and trust is going to be critical, especially in the public servant, public sector type context, because we think we need AI, because we have not enough public servants. But in order to flip it that way, we need to trust it, we need to have a way to communicate with the normal person.
Sepideh Pashami:And that's kind of where it stands, and also the type of the problem that comes to the picture, also often from a different nature. Often it's like certain administrative work potentially can be replaced by some sort of assistant tool or at least help them to facilitate the process lengthy process of doing certain procedures, and if AI can do that in a good way, reliable way, why not? But then the question is, as you all know, hallucinations and so on, that kind of plays a role and the question is can we detect those? I mean this opens up a lot of research question in this direction. That's, for example, can we detect a hallucination? Can we create explanations that kind of provide extra value, of why we are giving certain results and not the other ones, and so on and so forth.
Jesper Fredriksson:Would it be possible to do something like? I'm thinking about Skatteverket and all those government agencies? If they want to use LLMs, for example, it's today very difficult for them because they can't use any cloud resources for privacy reasons. And would it be possible to make that's a big project? And would it be possible to make that's a big project but like somehow that the government agencies can have some kind of shared repository, some way of that was, I think, in the report.
Henrik Göthberg:It's in the report and it's this common infrastructure and this is a rabbit hole.
Sepideh Pashami:I don't want to go into it because it's like how to build that, because central platform is great, but it it's like how to build that, Because central platform is great but it can be misunderstood how to do it.
Jesper Fredriksson:But it's in there. But it's one of those topics, right? Yeah, and it feels like the approach is maybe too large.
Henrik Göthberg:But you can flip it. It will be virtually impossible to do anything substantial for a municipality in Sweden. There needs to be something that they can. Is that impossible?
Sepideh Pashami:I don't like the word impossible.
Henrik Göthberg:No, I said virtually. I used the word virtually in front. It will be very tough for a municipality to do things.
Jesper Fredriksson:I see what's happening over there. She's already designing a solution. I like it. I like it.
Henrik Göthberg:With very small money, I could help.
Jesper Fredriksson:But you know what I mean.
Henrik Göthberg:What I'm saying is to have a fundamentally fragmented approach to AI across each municipality in Sweden. And then you need to think about you have Stockholm over here, but we need to have equal service with the smallest municipality in the smallest part and in the furthest way in Sweden. I mean, it's going to be super.
Jesper Fredriksson:So to find something that everybody can share in some way, I think is very I'm thinking about something small scale to start with and see how to do it in experiment. I'm all about something small scale to start with and see if it's possible.
Henrik Göthberg:How to do it in experiment. I'm more for it.
Sepideh Pashami:Yeah, I'm teaching in a course called AI for Executives and in that course actually in the latest sessions, more and more I'm seeing the presence of the different municipalities in those sessions. So I think this is changing.
Henrik Göthberg:I think it's changing. I think they're ready to get going, but we should help them. We should help them.
Jesper Fredriksson:I think, going back to my question about how can they use LLMs, for example, in a safe setting, about how can they use LLMs, for example, in a safe setting, I think there's a lot to be, a lot of simple things, simple building blocks that could be put in place. For example, the Mac Mini is a wonderful tool for running inference on LLMs.
Henrik Göthberg:It has a lot of memory and you can run them locally and you can stack a few of them on top of it and I'm with you here, because the tricky point is when someone tries to build a monolithic, central solution, I'm thinking more of a, of a, of a toolbox, a certified toolbox that you can use. So someone needs to put together a central toolbox. This is much more federated. This is much more data mesh thinking. Yeah, self-service infrastructure we're talking about. I'm thinking way more self-service infrastructure for the music. I do not believe in a central, monolithic platform. I'm going with you there and this is this is what scares me. When we say we're going to have a central infrastructure, what do you really mean? I, I mean self-service infra.
Jesper Fredriksson:Yeah, I think we need a few hackers from Reddit who can just put these boxes in place, on all the places.
Henrik Göthberg:But back to the question of the fundamentals. You know, applied AI research in enterprise and public sector, and I think this is the good timing when we sort of start digging into this topic. How do we think in this table around the distinction between research, applied AI research and AI engineering? So I think, because now this is, if I say we're going to do applied AI research in enterprise, now we're really starting to muddle the water potentially to do applied research in enterprise.
Sepideh Pashami:now we're really starting to muddle the water potentially so do you have a good definition of engineering versus research? I guess when I try to define a research project, I don't think about what are the needs today. I think about what will be the needs in five years. So I think that distinguishes between the two. Ai engineering they need to solve the problem of today, but what we do in AI research try to solve kind of bits in future.
Henrik Göthberg:So that would be. This is a good angle. So you're looking at a much longer angle that you don't have the force that this needs to be in production in a use case, in a business case next year.
Sepideh Pashami:No, we do. Actually we often, depending on where we apply for the funding, we sometimes have to define it in the term of a use case and then we also have to get an okay from a company that potentially this use case is interesting, have to get an okay from a company that potentially this use case is interesting, and of course the companies are involved because for them it's a small investment. If it's successful it's high pay and if it's not, then it's okay. But that's how we think about the future. For example I mentioned I'm interested right now in investigating a little bit on the robustness of machine learning, robustness with respect to the changes, unexpected or expected, and also to the adversarial attack, and also on explainability, and I wrote a proposal for a EU project. It's a very competitive course. Probably I will not get it it but potentially the idea is we deploy something like rich TRL 6 as a prototype, at least showcasing, in four different industrial cases.
Henrik Göthberg:Let's stop here a little bit, because you are using technical readiness level scale, trl, and then could you explain that scale from 0 to 9? What is 0,? What is 9? Technical readiness level scale, trl. Yes, and then could you explain that scale from zero to?
Sepideh Pashami:nine, what is zero? What is nine and what is six? It's hard to say all the levels?
Henrik Göthberg:No, say it in general, but I can say range yes, exactly, give a range yes.
Sepideh Pashami:So often, theoretical research goes from TRL one to three.
Henrik Göthberg:Theoretical research being what research?
Sepideh Pashami:goes from TRL 1 to 3. Theoretical research being what Theoretical research is basically the research that does not necessarily have a use case. It's fully theoretical. So you want just to advance the technology.
Henrik Göthberg:Grundforskning in Swedish. Yes, basic research, basic research yes, basic research and then TRL 1 to 3.
Sepideh Pashami:1 to 3, yes, research, basic research, basic research, and then trl one, two, three. One, two, three, yes, and then from four to five is basically prototype. But the prototype that is in experimental experimental setting prototype exactly.
Henrik Göthberg:Is this applied ai research?
Sepideh Pashami:yes, I think there is applied ai starts starts right yeah exactly. And then trl six and 7 is. Then you try to do a prototype, but in real world setting. So it's still a prototype, but not prototype. Let's call it like, tested in real world setting, what you have developed.
Henrik Göthberg:But I think this is very and then you go to production.
Henrik Göthberg:Because if we apply the TRL scale, we can get a more nuanced discussion on this. Because then you have theoretical research one, two, three. Then you have prototypes and pilots, but they're not ready for commercial production between four to six, seven, maybe six and then we get into what we probably call a POC in engineering style from six it's not ready, but it's closer to we really want to POC it in order to put that in the project for next budget year. Yes, and now we've got a quite nice spectrum theoretical research, applied research, you know, from four to six and then what we call engineering POCing.
Henrik Göthberg:Yeah, that sounds fair enough. So this is what. And then people can Google this TRL scale. You can read up on it on detail on each level and all this, but I liked your way of framing, sorting it. Anders has a favorite pet. Have you heard that one? Maybe you know, when we talk, what's the definition of research and what's the definition of engineering and I'm going to humor Anders, you know to try to say it like he does it he always says research like philosophically, research is the pursuit of knowledge.
Sepideh Pashami:Unknowns, basically yeah knowledge while unknowns.
Henrik Göthberg:basically, yeah, knowledge, the core pursuit is knowledge, while engineering is in the pursuit of product or in the pursuit of solving a practical problem, whilst and I think this goes very, very true for theoretical research, yeah, and now applied research is somewhere in between.
Sepideh Pashami:here and then it gets muddling and then of course I have a good example, but the reason that we need applied research is still you need to do some research to bring those technology to life and that's not obvious.
Henrik Göthberg:That's why there is still a need for doing applied but, but under underscore point of view that I I kind of. It grows on me and I I've been indoctrinated, let's call it like that. So if I take elon musk, elon musk is landing a rocket and gripping it with an arm. It's one of the most amazing technological feats of humankind. Is that research or engineering?
Sepideh Pashami:That's more of a result of the innovation? It's none of the two.
Jesper Fredriksson:I don't think it's engineering.
Henrik Göthberg:I think it's engineering too, because, if I follow under philosophical logic here, the pursuit is engineering to land a rocket and to grip it, and in order to solve that problem there are so many unknowns that you need to research. So you've done research with a clear engineering problem to solve, and that's the difference then with to theoretical research, in my opinion. So, so then we have the extremes right. So how would you? We now put applied research in the middle, because I think, applied research getting closer to engineering, you're having a concrete problem you want to solve yes, often, I mean sometimes, like, yeah, as I said, like maybe the difference is not on the.
Sepideh Pashami:You want to solve often the real challenge that exists in the either industry or in your society. So in that sense they are similar.
Henrik Göthberg:But then, um, you don't have the pressure that you need to do something that works good enough for now, so you have the flexibility of imagining what could be even better so that's kind of because, because I don't think that, because if I think it's so amazing, spacex right, and you know, because it's clearly I mean like they're blowing up rocket and they're amazed that it works. So in one way, there I think it's really they're not, they're safe, not to they're, they're they're set up to fail as well.
Jesper Fredriksson:Right to experiment yeah, I guess you could frame that problem of of landing a rocket and gripping it as a research problem exactly, exactly but it feels like what's happened at spacex is more like they're doing it iteratively and taking it one step at a time or hand in hand together, yeah, yeah, and, and using the research that's coming out of the daily work.
Henrik Göthberg:Okay, good, great. Now that leaves this is the perfect segue for my next question. Okay, how do we organize applied research or experimentation in enterprise in the best way? Research or experimentation in enterprise in the best way? And then we can sort of contrast. Now Vattenfall has an R&D setup, scania has an R&D setup, and it's clearly R&D on the side, and I would argue, spacex does not. I would argue that they are clearly doing R&D like this, but, if I understand it right, they're building two or three generations of rockets at the same time. So they have one rocket that is sort of closer to production and this is the Falcon that they're sending out right now in order to get all the you know, the Starlings up and stuff like that At the same time. They worked on the first Starship and they launched that, and they are already working on the next one. What the hell is research? What the hell is? You know, I don't know how they organize it in detail, but do we have any ideas? What is the way to do? R&d or applied research?
Sepideh Pashami:When I worked in this research project with companies, I often like to define the project in a way that not only I have some contact from the R&D in the team, but also I have some of the stakeholders of the end product in the team.
Henrik Göthberg:Yeah, this is great.
Sepideh Pashami:In this way it works much better than if I have either or.
Henrik Göthberg:But this tells us something about how to organize that if you're too far away from the real you are mitigating, being too far away from the real operation. Exactly that's what you're doing.
Sepideh Pashami:Yes, that's that's most of my successful project were in such a setup that they, we, I had a possibility of being closely working to the kind of pro product owner at the same time.
Henrik Göthberg:So, whether you like it or not, you gave it a response to how you should organize supplied AI research purely in R&D or closely.
Sepideh Pashami:And also like high-end research, like how putting these people together in the same room kind of makes wonders.
Jesper Fredriksson:Same room kind of makes wonders. So I, I think I still don't know the the difference between research and and engineering, but I, I really I think we're probably much closer to each other than one might think. It's, um, I like when you said five years into the future, I never think five years, but I try to think two years maximum, like one to two years, and and that's that tickles my brain all the time to to think about what's going to happen and always update the sort of the trajectory this is a good definition.
Henrik Göthberg:potentially to look at this what horizon are you working on? This is one way of separating it.
Jesper Fredriksson:And then I also liked when you talked about having the end users close to you. That's also in all my successful projects. That's always been the case working close to the end users or potentially being the end user myself. Those are the only two things that work, otherwise it will never work.
Henrik Göthberg:And this is the tricky point, because we have a tradition that we have an R&D department and I never worked in R&D in Vattenfall, michael Klingwald, who is in Dairdaks, he worked with me in the business and then he worked for several years after I moved somewhere else in the R&D department and of course, then they are working a little bit like the internal rights. That's what they're doing, what the file is, their internal rights, their material handling data, like this right, does it work? You know how close are they to real business problem? You know it is really, unfortunately, in my opinion, very human, related to you know who has to, who gets the relationship, who has the alliances, who works well together? Because I think the bottom line is, if you're not getting close to the real product or the real operations, it never gets beyond the prototype graveyard.
Sepideh Pashami:You usually solve the wrong problem, or even if you solve the problem, nobody's going to use it. That's another scenario.
Henrik Göthberg:So somewhere here there is an organizational lesson that whatever you're doing when you're trying to organize applied AI research, whether Jesper is highlighting it or is paid it, he's super close to the customer problem and the customer and the real reacher for the problem. I like that. So if we go into the future of applied AI research, okay, so we think, is it important? Is it getting more important or less important? That was my sort of follow-on question I started with. Is the way we are thinking about these topics now and you are thinking two years, you're thinking five years is it more or less important moving?
Jesper Fredriksson:forward. What do you say?
Sepideh Pashami:You mean moving forward in research.
Henrik Göthberg:No, no, no, I'm talking in general. How much attention, how much applied AI research budget should a normal company put? If they put 1% one year now, next year they should put five percent, why you know. So the way we are thinking about, oh, we need to really think about some applied ar research here. I can just do the if I do the setup.
Henrik Göthberg:It seems kind of important when I see the, the most valuable companies in the world, and how much money they spend on research in terms of the metas and the tech giants. I see, I see sahar and other ones working with the, the, the research arm of king. I see andes working at spotify ar research, or you know, you know all this. So the hyperscalers they have people that are clearly there to think not the next quarter, not the next year, but a little bit longer, and not just strategizing but actually testing stuff. So in that context, I'm not sure all companies are doing this I would not have a clue if I tried to do a spectrum of this, but the large enterprises are kind of doing it and now, with that sort of setup, is the importance to start thinking on these terms? Or even, maybe you need to put a research budget in there somewhere. Is that something that is applicable to more companies or that we should be more vigilant about? Is it growing in importance or not?
Sepideh Pashami:I think people are changing with technology so as a result also they would demand for higher quality services and nicer experiences and experiences in several dimensions, and this will kind of create a competition for the companies involved to kind of provide those services. And then, as a result, many of those services are AI-based, because just AI facilitates certain optimizations, certain conversations, even fastness in getting certain type of results and so on. So people will demand better services soon. They are starting already. After Chattopithi. We are seeing this that the expectation of answering something is getting lower, kind of the expectation goes higher of answering something.
Henrik Göthberg:It's getting lower, Kind of the expectation goes higher and this would infer by implication that we kind of need to experiment and research it in order to how the hell should we get it into our products.
Sepideh Pashami:Exactly so that's kind of the I think that's unavoidable.
Henrik Göthberg:What do you think?
Sepideh Pashami:And I think that's unavoidable.
Jesper Fredriksson:What do you think? I think I'm thinking about this gold field, where you can pick up gold, and I think it's a good time to be in the short to medium term, uh horizon of finding things, because there's so much out there right now, and that's why applied research, engineering, whatever you call it, but something that's that's directed towards where things are moving, because things are moving so fast, and just following that trajectory will be very beneficial. I think that's a that's a good path to be on. That's how I see it, uh, and then it feels like once we solve this thing, when we follow the trajectory, then we, then we will accelerate everything because, uh, base basic research will also follow with. Together with ai, you will be able to do research.
Henrik Göthberg:So just following this trajectory will solve many other things as well, I think and and there is a nice wayne gretzky supposedly said something that is super relevant to AI Go where the puck is going, don't go where the puck is. So, organize for the trajectory, learn for the trajectory, competence, develop for the trajectory, not for business. Right now, and because we're moving in such a fast pace now, if we are truly trying to think about the trajectory, it kind of implies back again, it's the same question. The trajectory is that the customers demand more efficiency, automation experience. The trajectory is that, whether you like it or not, we will work with agents in 10 years.
Henrik Göthberg:So you kind of need to experiment with it to figure out how you need to organize when you have an agent co-worker. If you haven't even touched it, if you haven't even done research on how that would work or experimented it, it will be a shock. This is the logic. If you organize for the trajectory and right now that seems like a good idea then that would call for applied AI research or engineering with a more future oriented take, maybe not necessarily solving the data table BI problem here and now, but something that you can't really use right now. But damn, you need to figure that out, you need to learn about it in order to be relevant. Yeah, I think that's the bottom line it in order to be relevant.
Henrik Göthberg:Yeah, I think that's the bottom line.
Jesper Fredriksson:Yeah, I mean, I think it's in a company, I think it's just about survival to be able to follow the trajectory, Otherwise you will not be sustainable.
Henrik Göthberg:And another way of talking about the trajectory have you ever used the word the productivity frontier? Is that a word in your vocabulary? No, so I don't know. I picked it up a couple of years ago and I I circled back to it, so we're always talking about there. You can research and google this, and or you know the productivity frontier, trying to sort of highlight you know what is the level of productivity in a certain operation, industry process at a certain time. And when we have technologies, that sort of the real paradigm shift is when we have a substantial cost reduction in doing something or speed increase in doing something. That is 10 hundred fold, and we have seen it a couple of times. Right, and now we are kind of in this level where you know what. We don't even need to talk about new business models, we need to simply talk about we will substantially do work in a different way. The productivity frontier will be somewhere else, and this is then going to be the tricky point. The trajectory is a little bit like where is the productivity frontier to be relevant? So the productivity frontier sort of kind of says you kind of need to be within these tolerances, otherwise you're obsolete, and that's the tricky one here. So the productivity frontier, I think, is what we're talking about with the trajectory discussion here. Interesting if I sort of what we are talking about with the trajectory discussion here. Interesting If I sort of wrap up the idea of the future of applied AI research.
Henrik Göthberg:Let's circle back a little bit to RISE and everything else. So I had a question in my head. I'm not sure if it works, but what is the ecosystem in Sweden around applied AI research? So we have RISE. We potentially have organizations starting to kind of build more research capabilities. So we say we're hinting they should research capabilities. So we say we are hinting they should. We have an AI commission coming out to say we need more spets. You know research in academia. So what's the evolution of the ecosystem for getting support with applied AI research? Should we do like RISE is doing and then you scale it up, or do we need something else? Do we need more boutique research firms, like should DERDAX pivot from a consulting firm to a boutique research firm? You know what's the ecosystem around applied AI research moving forward and think Vinnova, think AI Sweden, think RISE. You know whatever tools we have here.
Sepideh Pashami:Think AI, sweden, think RISE, whatever tools we have here For us. We get our assignment in two different ways often. So some we call it direct assignments, that the companies come to us and directly ask us to do a certain type of assignment. They are often very clear what they want and also which time frame. And then we have another type of assignments that are kind of funded by public funding agencies. We write proposals together with startups, bigger companies at the same time and some universities, and this is often we submit it to some funding agencies, for example Winova.
Sepideh Pashami:But it's not just Winova, there are others like Energy, mindy, there are Forte. There are funding agencies that they have a team, and some of them have a team and allow this kind of applied research, and some of the smaller universities also follow the same trend. I would say that they apply for these funding possibilities. And then, of course, these funding agencies come with a team, in many cases with a team. In many cases, for example, vinnova might have a strategic program for strengthening the vehicle companies, like FFI program. They have also for other sectors. Then you try to formulate your problem in that setting and then potentially with the right partner, that kind of create a chain of partners that you need to solve that problem, and research competencies also, and then you try to solve the problem. So that's how it works right now. There has been discussions that they would like to unify these different smaller agencies to one bigger decision-making process.
Henrik Göthberg:There's another angle on this Same question, on the ecosystem, and it's a little bit like the role universities are playing versus the role RISE is playing versus the role like AI Sweden is playing or consultancies are playing. But let's make it simple. Make it RISE universities and industry consultancies. You know, are the roles well set up? Are we clear? Are we overlapping? Can we do it better?
Sepideh Pashami:Because sometimes it seems a little bit like muddy it's a bit overlapping, but at the same time, I see at least one clear role for each of them. So, for example, I see AI Sweden as a very good networking place. I know they are doing certain type of research, but in a just particular research. That is kind of strategic maybe, but I see them as a very good networking that we need. I think we need to be connecting body. We need a connecting body between even more than what they are doing. I would say we needed to connect academia and industry and smaller sectors, public sectors and even government.
Henrik Göthberg:SL. So it's a connecting role, not doing the research per se, but connecting and federating between more… F I mean they are doing some research, but I don't think they can scale Like that's just… those research.
Sepideh Pashami:I know some of my colleagues are there, so I'm… I love their research.
Henrik Göthberg:SL. It's simply the scale there. So I'm I love their research. It's simply the scale comparison, but they will be exactly.
Sepideh Pashami:So the scale would be the question. I think maybe for that particular topic they will be excellent, but then of course they cannot cover all the possible topic. We need AI in the society, so that's that's quite a huge demand. And then so I see their networking role as a bigger one. We need it, we need it. As you said, today I met Jesper. Why we should have met?
Henrik Göthberg:years ago. The number one network is AI After Work podcast. Exactly.
Sepideh Pashami:So I think that's kind of the things I would like to see more often. And then universities' main role is education, so we need next generation AI engineers, ai researchers, ai developers, not only AI programmers, software developers. I mean we need all of this chain and if we don't have it, then we don't. And then when it comes to RISE, rise is there to support Swedish government, swedish industry and also Swedish public sector. I would say so to basically help Sweden stay high-end research strong and so on in Europe and in the world, I would say.
Jesper Fredriksson:And to continue on the question of the future of applied research. Does RISE see the need to change how they work with with a sort of new, new technology, new gen, ei, everything moving fast, or is is that still invariant? Does do we need to change? Does rice see the need to change because of, like new um?
Sepideh Pashami:because we are working on research, so we also have research on like language processing and so on. So it's like yeah, but not.
Jesper Fredriksson:You don't need to change like the way you work, because I think that a lot of organizations like volvo needs to change, and you were talking about universities. I think the role of universities might also change. When, when you an AI tutor, for example, do you see the need in RISE, to change how you work towards companies or how you do your research?
Sepideh Pashami:What do you think will happen? I like your question. I have been seeing a transition from developing new AI system and utilizing the existing one. And we are seeing this at RISE as well that we have to work in both.
Sepideh Pashami:And actually there is a lot of demand for utilizing the existing ones for different application area more than the ones that we would like to develop new technology at the moment. But I think in order for us to stay in a good level of basic research we need to have both. That's my personal understanding. But then also there was one more factor about consultancy. I think what we are doing is often they are prototypes, so they needed some other bodies to take those ideas and kind of take steps toward a product, and I think here the role of the consultancy companies or the bigger companies comes to the picture and without them we will not be able to deploy anything, basically.
Sepideh Pashami:So I think, um and and as, as we were discussing, there is also overlap between them. So also universities. They do research, some of them they are doing applied research, and I don't see any problem there. So that's great.
Jesper Fredriksson:We need more. Why not? Yeah, I think what I'm seeing in enterprises and companies in general, the main thing that needs to happen is that people need to be more able to change, change faster, change how we work, be prepared for us, as Henrik was saying all of a sudden, maybe your new co-worker is an AI agent. We have to be prepared for all the change that's going to come. So this adaptability in the humans that's going to come. So we, so this adaptability in in the humans, that's, uh, that's going to be a key trait, I think. Um, I'd also like to touch on, on, like the universities you mentioned universities doing basic research and doing education. Um, you're still in halmstad University. What do you think about the role of Halmstad University, for example, into the future? How will they change?
Sepideh Pashami:I mean as teachers, we think a lot after, especially like the advancement in the generative models. We have been thinking a lot about like how to utilize the, for example, generative models and. Chachipiti and many others in our education system, because the student is going to use it.
Jesper Fredriksson:Yeah exactly.
Henrik Göthberg:Yes.
Sepideh Pashami:It will be stupid of us to just say imagine they should not tell them don't use this.
Jesper Fredriksson:Like this doesn't work.
Sepideh Pashami:So we have been thinking how we can utilize it at the same time because, like all tools, if you use it in the right way, it can teach you a lot, but if you use it in the wrong way, it will stop you from learning. So I think that's what we need to teach and adapt in education for this generation and also examination has become more and more complex. For example, we give assignment coding assignment. Assignments are simple because they purposefully designed simple for teaching purposes.
Sepideh Pashami:And of course, tools like Copilot could solve them like that incorrectly, and so it's a bit of a challenge. I'm not saying it's easy, but we have been trying to adapt. For example, we added a prompt engineering exercise in my course using these LLMs, so we are trying to just teach them things on how they should use the system and some of the written assignments turn to presentations and so on, just to make sure they understood what they are presenting.
Henrik Göthberg:So there was something about the future of applied air research and the different roles we take. Now we're getting to the very ending stretch of this um, marathon, marathon to our marathon and let's call the last stretch looking ahead to agGI.
Henrik Göthberg:So me and Anders always end with the last final question with the spectrum. But since me and you are both kind of nerding down on agent ideas and all that, there is a couple of interesting conversations that I couldn't help that maybe before we take the end question I at least want to have one or two questions on. Maybe before we take the end question I at least want to have one or two questions on. So we are talking now moving ahead towards AGI and we might think about we might start building gigantic workflows that in the near future.
Henrik Göthberg:In the end you are quite close or you get to mimicking a co-worker in a team and I find that a quite interesting topic. Simply, if you take the human-machine interaction, we have a whole field in KTH. We have an whole department talking about human-machine interaction and I don't think we have maybe I'm wrong, but I don't feel that we have ever had a human-machine interaction where we have a principal-agent relationship with a machine. I think this is different. I think to have the machine as a co-worker is fundamentally different than what we have today.
Henrik Göthberg:Do you agree?
Jesper Fredriksson:Yeah, I mean human-machine interaction is basically like how do you design a UI if you? Want to dumb it down really much, but this is something else.
Henrik Göthberg:And when me and Mikael you met, mikael, we have been going deep down the rabbit hole. You saw our presentation on agency and again thinking. And in organizational behavior and organizational research there is a deeply researched problem called the principal-agent problem. Basically, how do we set up the boss versus the coworkers and how we control and steer them? And in that research terminology we are now referring to a principal-agent relationship. Someone is the boss and someone is the coworker and this is now, in my opinion, the epicenter of this new type of human machine interaction. Do you follow me? Could we sort of elaborate? What does that mean? Have you thought about this?
Sepideh Pashami:have you thought about this? Not extensively. So I like the idea of having assistant co-worker, for example, when I want to explain the role of the machine learning in a hospital, let's say, if I want to explain the role of the machine learning in a hospital, let's say, if I want to describe it to a nurse, I will imagine it or I will describe it. You can imagine, this is the new assistant you have and this assistant will tell you this is the diagnosis and at first, of course, it's a new colleague. You, you might not trust it, and then and you you should be like that, rightfully so and after a while, maybe observing certain cases that you're agreeing with, then you trust, and then suddenly there is a case you don't agree with and then you again doubt it and that if this is a good colleague or not. So that's how I describe the relationship that potentially can happen, I mean, even if it's not like a with a human body it will play that role.
Henrik Göthberg:But but let me sharpen and make this a little bit more futuristic. So there's just a level one assistant. Yes, we know the solution and we are asking the assistant for a direct approach and we don't really give over agency in any way. So the first level of assistant we ask them and we get a response. There is no agency on the other side. So when we say agent on the full level, we are actually giving them the problem and then, based on the problem, they break down that problem and they start executing on tasks. So where I'm going with this, when we have a co-worker, is that we are releasing. The difference to me, an assistant, to a co-worker, is that you have released agency, and I think you know. So what you said is level one yes, tick in the box. Take it one step further. When you, you know it's not only an assistant, you're going to go away and let them do your job, not recommend you what to do. I think that's when my head starts spinning. How you know, do you follow?
Sepideh Pashami:Do you follow the difference? You're talking about this basically when the AI system can have a human level of intelligence in a sense that Not necessarily, but within a frame.
Henrik Göthberg:They can solve a problem and do the work for you.
Sepideh Pashami:So I would phrase it as if they don't make mistakes that seem stupid to humans. Is that good?
Henrik Göthberg:Okay, I like the way. In order for us to be able to give up agency to someone else, even how narrow it is, they need to provide a service on the level of a human. The first clear example we've seen is autonomous driving. Exactly, we're giving up agency to the car to drive from A to B and in one way it can do it pretty good now, but we're not ready now yet to regulate it. I can imagine other areas, which is not so critical, where we actually can give away agency. Maybe not in healthcare and stuff like that.
Jesper Fredriksson:But I think this is what we're talking about. I guess what you're saying is, when you're in the car you're not saying take me to uh, to mcdonald's. You're saying I want to eat.
Henrik Göthberg:Is that is that sort of I mean, like even with the car, I think it's enough, just take me somewhere.
Henrik Göthberg:If even the agency even the agency to drive from stockholm to malmo to me without me caring. I think that that that is quite a bit of agency. To feed me on the way, well, that would be even cooler. But I think the autonomous driving example is quite strong, as it is in relation to agency to decide on how to take me to Gothenburg and to Malmö and then to how to maneuver traffic. But so I'm saying it's more like I don't know what I'm saying, but I think it's this trajectory we're talking about here and how to deal with this giving up agency to a machine. I think that is a very interesting question.
Jesper Fredriksson:I thought I understood what you meant and the way I interpreted it was I probably what you meant. The way I interpreted it was I probably misunderstood you. But if you're seeing it sort of when the AI agent is performing a task like driving from A to B, that's some freedom. You can choose any path.
Henrik Göthberg:There's some agency here for sure.
Jesper Fredriksson:But the next level is more like you're sitting down in the car and you don't know what you want, but the car already knows that you're hungry. You should go to mcdonald's first and then you want to. You want to get some entertainment, maybe you want to go and and you just sit in the car and the car takes you to places and tells you to get out.
Sepideh Pashami:That's uh that's what I thought that you meant, and then play, play music.
Henrik Göthberg:But you're taking it further, because I was somewhere, you know, beyond just simply keeping the lane. You know, I think this is a trajectory and I don't think we even need to go all the way to the sort of hypothetical approach that they are thinking for me. I think the simple task is that please could you do the triage at the thinking for me. I think the simple task is that please could you do the triage at the hospital for me. I want you, as a robot, to do the triage for me. You have the agency to do the triage for me.
Henrik Göthberg:It's an interesting thought, right, because you can ask the questions, you can get the video, you can look at that and then you need to do the decision should you sit down or should I send you to the operation? It's more than just instructing. It's more than just giving the core. You know this could potentially be dangerous. It's making a decided action in relation to something. This is quite extreme example, but I'm so. It doesn't need to be so difficult, it doesn't need to be so big, but it's actually making execution on topics. I think that's when we are really starting to talk about agency yeah, well, I mean, that will come it will come right.
Henrik Göthberg:But are we ready? But? But how do we think about that? What are the big? Uh, you know that that we, we need framework for how to think about this.
Jesper Fredriksson:Probably, yeah yeah, I think if it's just me, then then I think I know the framework. But if I'm thinking about volvo, if releasing agents in volvo, that's going to be an interesting experience to see what will happen, because it's you want it to be aligned with the sort of company goals and what will happen then if you have something at one department and then there's another department and maybe there's conflicts between the departments and they start fighting or doing some, something that's contradicts each other. That's one thing, and another thing is, let's say that that I have my personal assistant, that's my agent. I would be tempted to take that into the organization. It's like bring your own AI kind of problem.
Jesper Fredriksson:What happens then if you take your own personal AI and move it into another setting which is not optimized for the organization?
Henrik Göthberg:I don't have time for this meeting. I sent my avatar.
Sepideh Pashami:And then I will take the summary at the end.
Henrik Göthberg:But I think the way me and Mikael have been thinking about that is the fundamental feat of aligning human and artificial agency.
Henrik Göthberg:And now comes to the core topic.
Henrik Göthberg:This is part of the thesis that we need to do a different panel discussion on, because mine and Mika's view is this that if this is going to work in a good way, where we can input agents into this, we most likely need to go to a more agent-based organizational model.
Henrik Göthberg:Agent-based organizational model. So the way the organizational objectives are quite clear with agency and how they interact with each other, because then I can put an agent in there If I have a very matrix organization with very conflicting objectives and say the simple thing that you can technically create an agent that spans two or three teams, easy, you know, in terms of if you're organized stupidly, right. So if you have functional stupidity, if you have the way of organized a company or a department and then you place an agent in the middle of there that it doesn't, it's not, it's not aligned, or or actually it is probably aligned with the work, but the organizational mandates are not organized accordingly, I think this is problematic, right yeah, that's going to be so, so, so, so the whole thing, when you put agents in the mix and you understand how agency of artificial agency works.
Henrik Göthberg:Mine and Mikkel's aha moment has been we need to sort out human agency at the same time we need, we need to, or, let's say, team agency. The tricky point is when you have an agent that goes beyond one team, then it's gonna and they have conflicting. Do you see what I'm talking about?
Jesper Fredriksson:yeah, and I like when you talk about human agency, but because it sounds a little bit like hiring a co-worker so it's like that's maybe not so complicated. I guess the the thing is that this might have superpowers and it might not respect any any rules, that's why or also like it might become a power game, exactly, yes, whoever has the?
Sepideh Pashami:better agent kind of can solve the problem and then the other ones become irrelevant.
Henrik Göthberg:Yeah, in one way it could be very simple, like you said, and I think you're right. But then there is another step two on this. So typically, as a co-worker, you work in one team and you're supposed to work in that team, but whether you like it or not, you probably have a couple of different feedback loops that are human that you're calling to colleagues in other departments. I mean, like I'm doing the marketing planning now I kind of need to talk to the sales guys and the product guys on what the marketing planning is going to be all about. But my work is marketing right.
Henrik Göthberg:So what if those feedback loops that we are now typically solving through human capital is not set up properly? Then you have an isolated marketing assistant co-worker that doesn't have the human feedback loops that a human co-worker would have. The human co-worker would solve it by simply going over to his friend and talk to hey man, what do you want to sell? I'm going to market something here. So it sort of puts an extra stress on okay, great, co-worker, but we now need to encode the feedback loops, all this.
Jesper Fredriksson:I mean, they will maybe have one-on-ones as well. Yeah, with another agent. Maybe one agent will call the other agent and give the feedback. Maybe it's not so problematic.
Henrik Göthberg:No, I'm not sure. I'm not sure where the problem will lie. I think it's just a thought experiment that is actually not so far away. I don't think it's that far away.
Sepideh Pashami:In that regard, I would like to say that, for example, when Chachapiti came in a few years ago and the way it came in and became part of the society was interesting, so that's, I would say, what was the closest thing that we have experienced to this artificial general intelligence that you are talking about and the fact that it has been released into a large population create an initially worried at the same time excitement at the same time.
Sepideh Pashami:Everybody wanted to have it but at the same time available to everybody. So that was kind of a nice experience and it took a bit to settle I would say like a couple of months-ish to kind of people figure out what is this, how they can use it, can they have accounts and so on. But now that we are looking at it, it's not far from that moment. It's so much part of the everyday work of so many people already Already, and that's what I'm actually, that's what seems to be. We are more adaptive than we think we are. I understand that initial days was surprising and stressful and worrying and so on, yeah, but we see that we are now adapted to it. We are, we accepted it as part of this and maybe that's the solution.
Henrik Göthberg:So we can theorize about this as much as we want, but it's when we start adopting agents. There will be more junior agents. The agency will not be there from day one, so we will probably learn by doing again. It's the only way to figure that out.
Sepideh Pashami:I think so.
Henrik Göthberg:I think so too. Final question I can't do the same beautiful setup as Andres is doing but we are always finding the last question, talking about AGI, and we always set it up like a spectrum. So if you imagine we're getting to the point where we have some sort of AGI, we're not usually saying superintelligence, we're saying AGI. You have to adapt it now to ASI, I think. What is ASI then?
Jesper Fredriksson:Superintelligence.
Henrik Göthberg:Superintelligence, yeah, but we have this. And then you can think about a spectrum where it's sort of a dystopian view the terminator view it all went to shit, you know, the AIs took over versus the utopian view where we kind of live in a world of abundance. We can work if we want to, we can pursue the work we want for fun. If we want to, we can pursue the work we want for fun. And then, basically, how do you see the future? Where do you put yourself on the spectrum and what's your thinking about? You know what is the trajectory we are most likely to see.
Sepideh Pashami:How do you see this? This is yeah, this is an interesting question. So when it comes to one end which is, like Terminator, dystopian, I don't think we, as human, will let AI system to reach that point. I think we, as humans, we will create safeguards along the way that doesn't let that to happen. Hopefully, we have been like that through generations, if we look at it.
Jesper Fredriksson:We're quite resilient as a species.
Sepideh Pashami:So in that sense I don't think that I'm not on that end, but I don't think it will be totally on the other end either, but I'm more toward that utopian side. I'm more toward the utopian side, but not the end that you described. The thing is still there will be inequalities because depending on who has this AGI and how it will use it, then we might end up into something better, something good or something better, and that's that's. We remain to see who has it first. I would say it's interesting.
Henrik Göthberg:Sverker Jansson, I think, made a really cool answer where he, when he said, oh, we will have both. So we look at the world right, we have poverty, we have richness, and and why wouldn't we have some dystopian aspects, or so you know? I find that interesting. I think you made your way of framing it with it. There's not going to be completely perfect or completely bad. More on that side, I think it's very interesting one. What about you? You, uh, have you? You asked that? You answered that question once already, I guess yeah, you changed your mind.
Jesper Fredriksson:Yeah, exactly have you changed your mind, yet uh, yeah, I think last time I said, uh, I'm concerned about, uh, that there is no, no real movement towards what's going to happen, um, with, like unemployment, people be people whose jobs are being replaced. Um, there's been talks about UBI, universal basic income and stuff, but I don't see any real movement in that direction. So that's that's what I think is one of the scary parts another.
Henrik Göthberg:There's so many unanswered topics. How, how will we sort things out? I?
Jesper Fredriksson:think another, another dystopian scenario is if you look at what happened with social networks, it's like people find themselves locked in the social network world. And when you have an AI that can always give me nice answers, tell me how clever I am, which Chachib T does in a very nice way. You always feel like oh. I asked a good question so it's like it might be more interesting to talk to an AI than to talk to another human person.
Sepideh Pashami:That's also another scary future scenario but on the positive end, I always like to have my car as my best friend. Hopefully, which car do?
Jesper Fredriksson:you have and I always like to have my car as my best friend.
Sepideh Pashami:Hopefully have your car as your best friend. Which car do?
Jesper Fredriksson:you have, I have a Volvo, hopefully Excellent, I love it.
Henrik Göthberg:But can you imagine which car would be your best friend? It's a Volvo.
Sepideh Pashami:No, hopefully Volvo would be there first Sepede. Fantastic, paul, thank you. What a great episode. Thank you so much.
Henrik Göthberg:Hopefully Volvo will be there first. So, Peter, fantastic pod. Thank you. What a great episode. Thank you so much.
Sepideh Pashami:Thank you very much for having me. Thank you so much. Thank you, I enjoyed the discussion, yeah.
